Chipper Cash Thwarts Fraudsters in Real-Time with Pinecone Vector Search

Title

Chipper Cash Thwarts Fraudsters in Real-Time with Pinecone Vector Search

Content

Chipper Cash, an African fintech serving more than 5 million customers across seven countries, cut new-user identity verification latency from up to 20 minutes to under 2 seconds and reduced fraudulent sign-ups by 10x by replacing a third-party KYC service with a Pinecone-backed in-house facial similarity system deployed in Uganda and other markets. "End-to-end latencies for the entire system dropped from up to 20 minutes previously to less than 2 seconds now, with Pinecone doing the search in under 200ms," said Samee Zahid, Director of Engineering at Chipper Cash, who led the build. He noted the rationale for in-house development: "We have a high bar in terms of security and latency for our users. Many third-party solutions don"t meet our requirements, so we typically opt to build or host in-house. Pinecone proved to deliver so much value — with reduced overhead and ultra-low latencies at scale — we didn"t need to do much convincing to move forward." The shift was driven by a third-party KYC bottleneck that gave fraudsters a 20-minute window to exploit paid new-user promos such as "Buy stocks worth at least $2 before the end of the week. Get a ₦500 reward!" Across a six-month period, fraudulent sign-ups — duplicates plus users with fake or stolen government IDs — consumed 16% of Chipper Cash"s promo budget, forcing the company to end promos early and forfeit legitimate user acquisition. The deployment was completed in under a month. The engineering team first built a proof of concept using an open-source vector database and a ConvNet model that converted user selfies into embeddings, then concluded that scaling and managing the open-source stack to Chipper Cash"s latency and growth targets would be too operationally heavy. They adopted Pinecone as a managed alternative. The technical stack combines a ConvNet facial embedding model, Snowflake as the data warehouse, the in-house Facial Similarity Service (FSS) that generates and queries embeddings, and Pinecone as the vector database. For each new sign-up, FSS embeds the user"s selfie, queries Pinecone for the three most similar stored embeddings, and passes those matches to the Chipper Backend for likelihood scoring and metadata enrichment. Pinecone supports billions of vectors and returns the top three matches in under 200ms. Chipper Cash"s in-house facial verification system now achieves over 95% accuracy in less than 2 seconds end-to-end, with Pinecone contributing sub-200ms vector search. The fraud reduction of 10x has been observed across all markets, with Uganda as the most-cited example in the published results, and promo budgets now reach legitimate users at the intended rate. Motivated by this success, Zahid"s team is now exploring additional use cases for AI and the Pinecone vector database across the business.